UCI datasets.

The majority of existing clustering algorithms, including those algorithms that focus on boundary detection, seldom account for the reasonableness and genuineness of boundaries, consequently, it is difficult to obtain well-defined boundary in clustering-based regional division. A novel boundary search Clustering algorithm integrating Direction Centrality with the Distance of K-nearest-neighbor (DKCDC) is proposed, which is capable of achieving well-defined regional boundaries, to resolve the challenges mentioned above. Firstly, the preliminary boundary of clusters are established on the basis of boundary points and initial cluster labels obtained by the Clustering algorithm using the local Direction Centrality (CDC). Secondly, all the boundary points are further processed and discriminated, to detect noise points concealed within the boundaries, which provides the essential basis for achieving more genuine and reliable cluster boundaries and regional identification. In this process, a fusion strategy is adopted, to subdivide the boundary points into true boundaries and false boundaries by combining voting method and distance metric. Thirdly, a regional division result with well-defined boundary is obtained by DKCDC. In the end, by distinguishing genuine from false boundaries using fusion strategy, DKCDC enhances regional boundary demarcation. Experiments on synthetic and UCI datasets show DKCDC improves silhouette coefficient by at least s4.88% over CDC, K-Means, DBSCAN, OPTICS and HDBSCAN, indicating its broad potential for applications in clustering-based regional division.

现有绝大多数聚类算法（包括聚焦边界检测的算法）均未考量边界的合理性与真实性，因此基于聚类的区域划分难以获得清晰明确的边界。为此，本文提出一种融合方向中心性与K近邻距离的边界搜索聚类算法（Direction Centrality with the Distance of K-nearest-neighbor, DKCDC），可实现清晰的区域边界划分，以解决上述难题。首先，基于采用局部方向中心性（local Direction Centrality, CDC）的聚类算法得到的边界点与初始簇标签，构建簇的初步边界。其次，对所有边界点进行进一步处理与甄别，以检测隐藏在边界内的噪声点，为获得更真实可靠的簇边界与区域识别结果提供必要基础。在此过程中，采用融合策略，结合投票法与距离度量将边界点划分为真实边界与伪边界两类。第三，通过DKCDC即可得到边界清晰的区域划分结果。最终，DKCDC通过融合策略区分真实边界与伪边界，优化了区域边界的划定效果。在合成数据集与UCI数据集上的实验结果表明，相较于CDC、K-Means、DBSCAN、OPTICS及HDBSCAN，DKCDC的轮廓系数至少提升了4.88%，显示出其在基于聚类的区域划分任务中具有广阔的应用潜力。